Yangchun Ye, Tian Qin, Xuan Guo, Wenjun Wang, Qiyao Peng, Tianpeng Li · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840804
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Large Language Models (LLMs) have advanced text-attributed graph (TAG) learning in recent years, leading to strong performance in Graph-LLM systems. However, these gains may partially arise from target-correlated parametric priors memorized during LLM pre-training rather than genuine graph reasoning. This issue is particularly critical in TAGs, where graph topology can propagate memory-triggered signals beyond individual nodes. Once encoded as node representations, memorized priors may spread through neighborhood aggregation, introducing topology-level bias amplification absent in conventional NLP tasks. Despite its importance, the interaction between LLM memorization and graph message passing remains largely unexplored. In this paper, we present a leakage analysis framework guided by Structural Causal Models (SCMs) for Graph-LLMs. Specifically, we investigate how prompt interventions influence downstream graph prediction and examine whether entity-level anonymization suppresses memory-triggering cues. We further construct globally consistent entity-anonymized graph benchmarks that preserve graph structure while masking identity-sensitive entities. Experiments on citation and person relational networks across multiple embedding models and graph backbones reveal three findings: (1) Target-related meta-information can act as shortcut signals when highly correlated with downstream labels; (2) Graph message passing may further propagate memory-induced bias across neighboring nodes; and (3) Graph-LLM performance drops notably after entity anonymization. These findings suggest that current Graph-LLM benchmarks may overestimate generalization ability when memorized priors are not properly controlled. The code and anonymized datasets are publicly available at: https://github.com/mrjc12/GL-DataLeakage.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.